2026-10-12·Daily digest

🔬 Research Pulse

Daily Digest

October 12, 2026

🤖 AI

🤖 Agents

1. X-AgentRed: Red-Teaming Personal Agents in Evolving Multi-Service Ecosystems

Authors: Yeonsung Jung, Zhiyuan Wang, Zifan Wang... Published: 2026-10-09 | Citations: 0 arXiv | PDF

Research Question: How can red-teaming systematically discover novel compositional harms in multi-service personal agent ecosystems, rather than being biased toward familiar tool combinations learned during training?

Summary: X-AgentRed is a self-evolving red-teaming framework that builds a directed graph of tool dependencies (where one tool's effects satisfy another's preconditions) to discover compositional harms in personal agents spanning multiple services. It uses execution feedback to diversify paths via tool substitution and refine blocked preconditions, achieving substantially higher harm discovery and tool coverage than prior methods.

Key Results: X-AgentRed achieves 1.77x the harmful path rate of STAC baseline across four frontier agentic models, and yields 2.99x the tool coverage on average, demonstrating both higher attack success and broader exploration of the tool space.

Key Findings:

Graph-based path sampling discovers 1.77x more harmful paths than STAC across four frontier agentic models
The approach achieves 2.99x tool coverage on average, indicating LLM-only red-teaming significantly under-explores the tool space
Execution feedback (tool substitution + precondition refinement) is critical for breaking past training-data bias in attack generation

Technical Novelty: Construction of a directed tool-dependency graph where edges represent effect-to-precondition satisfaction between tools, combined with execution-feedback-driven self-improvement via tool substitution (for diversification) and precondition refinement (for unblocking failed paths). Prior work either prompts LLMs directly or decomposes human-specified harmful goals, both of which inherit LLM training biases.

What's New: First red-teaming approach that constructs harmful paths from structural tool-dependency graphs rather than LLM free-generation or top-down decomposition of harm goals, enabling discovery of novel cross-service interactions that LLMs would not propose from priors alone.

Extension Opportunities:

Apply the tool-dependency graph construction to defensive side: use discovered harmful paths to auto-generate guardrails or precondition checks that block compositional attacks before execution
Extend from harm discovery to harm severity scoring by weighting graph edges with blast radius (data sensitivity, reversibility, number of affected services) to prioritize patching
Adapt the self-evolving framework to red-team specific real-world agent deployments (e.g., Claude's MCP tool ecosystem, OpenAI's GPTs) by seeding with their actual tool manifests

Replicability: No code/data availability is mentioned in the abstract. Reproduction would require access to four frontier agentic models (likely GPT-4-class, Claude, Gemini tier) plus a sandboxed multi-service tool environment — moderate compute for the models but significant engineering lift for the simulation infrastructure.

Research Gaps:

No evaluation of whether discovered harmful paths transfer to real production agent deployments versus the paper's simulated environment
Unclear how the framework handles tools with ambiguous or probabilistic effects (e.g., LLM-generated content tools) where precondition/effect modeling is itself uncertain

👁️ Vision

1. Images Talk Back: Guiding MMDiTs by Suppressing Image-to-Text Feedback

Authors: Yucheng Yang, Xiaoxiao Sun, Shusheng Yang... Published: 2026-10-09 | Citations: 0 arXiv | PDF

Research Question: In multimodal diffusion transformers (MMDiTs), does image-to-text (I2T) feedback during joint attention help correct deviations from the text prompt, or does it actually preserve errors already present in the evolving image — and can suppressing this feedback improve generation quality?

Summary: This paper identifies that image-to-text feedback in MMDiT joint attention tends to lock in whatever is already in the evolving image — including errors — rather than correct it toward the prompt. The authors propose Semantic Feedback Guidance (SFG) and its in-place variant iSFG, which suppress I2T feedback as a guidance signal, yielding measurable improvements on GenEval, HPSv2.1, and VBench across image and video generation backbones with little or no extra compute.

Key Results: The authors demonstrate that I2T feedback preserves image errors rather than correcting them, verified via transplantation experiments (transplanting I2T messages rescues failures or disrupts successes). Their proposed Semantic Feedback Guidance (SFG) and in-place iSFG show concrete gains: on FLUX, SFG+CFG raises GenEval from 0.6661 to 0.7479 (+0.0818); on SD3.5M, iSFG adds +1.33 HPSv2.1 over CFG at zero extra NFE. SFG requires only ~1.2× the base NFE (without CFG) by running during the first 20% of sampling steps. The method also transfers to text-to-video and image-to-video, improving VBench scores across three backbones.

Key Findings:

I2T feedback in MMDiTs preserves image state (including errors) rather than correcting toward the prompt, demonstrated by transplantation experiments that rescue failures and disrupt successes
Suppressing I2T influence improves generation while symmetric 'opposite' controls degrade it, confirming directional causality
Applying SFG only during the first 20% of sampling steps captures most of the benefit; iSFG achieves gains with zero extra NFE and SFG adds only ~1.2× NFE overhead
The technique generalizes beyond image generation to text-to-video and image-to-video, improving VBench

Technical Novelty: Prior work on diffusion guidance (CFG, autoguidance, perturbed attention guidance) manipulates text-to-image conditioning or model noise predictions. This paper is the first to isolate and manipulate the I2T feedback pathway within MMDiT joint attention as a guidance signal, introducing an auxiliary branch that strengthens I2T/weakens T2I to serve as a 'negative' direction, plus an in-place variant (iSFG) that achieves similar gains at zero extra NFE.

What's New: First work to treat the I2T attention pathway specifically as a lever for guidance rather than treating MMDiT joint attention as a monolithic block; introduces an auxiliary-branch guidance scheme (SFG) and a no-overhead in-place variant (iSFG) that cleanly separate 'strengthen feedback / weaken conditioning' as a negative direction to steer away from.

Extension Opportunities:

Apply SFG-style attention-flow guidance to other multimodal transformer architectures beyond diffusion (e.g., autoregressive image models like Chameleon or unified multimodal LLMs) where cross-modal feedback loops exist
Develop adaptive scheduling that dynamically decides when to apply I2T suppression per-token or per-region based on prompt-image alignment signals, rather than fixed 20% schedule
Extend to editing/inpainting tasks where selectively preserving vs suppressing I2T feedback could enable finer control over which image regions update toward the prompt

Replicability: Code is publicly available at https://github.com/yyc1215lll-creator/SFG. Reproduction requires inference-only compute on FLUX, SD3.5M, and a third backbone — no training needed. SFG costs ~1.2× base NFE; iSFG is free. A single high-end GPU (e.g., A100/H100) sufficient for sampling; GenEval and HPSv2.1 benchmark evaluation adds standard eval overhead.

Research Gaps:

Mechanistic understanding of why I2T feedback preserves errors — the paper is largely empirical and does not fully characterize what properties of the attention weights cause the preservation effect
Interaction with fine-grained spatial control (e.g., layout conditioning, ControlNet-style signals) and whether I2T suppression could hurt in cases where preserving image structure is desirable, is unexplored

2. InstantFusion: A Shared Latent Interface for Heterogeneous Image Generators

Authors: Xuanlang Dai, Jiazi Bu, Yujie Zhou... Published: 2026-10-09 | Citations: 0 arXiv | PDF

Research Question: How can heterogeneous image generators (diffusion/flow models with different architectures and latent representations) share intermediate denoising states without requiring pairwise mappings between every model combination?

Summary: InstantFusion introduces a lightweight shared latent interface that lets heterogeneous diffusion/flow image generators hand off intermediate denoising states to each other through model-specific autoencoders projecting into a common latent space. The design uses two training objectives—noise-level state alignment and local velocity synchronization—avoiding the quadratic blowup of pairwise model mappings. It enables acceleration, multi-preference composition, cross-architecture distillation, and cross-model composition with a single trained interface.

Key Results: The authors empirically demonstrate two findings: (1) independently trained generators exhibit similar latent spatial organization, and (2) directly handing off intermediate states between unaligned models preserves coarse visual content. They validate InstantFusion across four tasks—acceleration, multi-preference composition, cross-architecture on-policy distillation, and cross-model composition—reporting 'competitive performance.' The abstract does not cite specific quantitative benchmarks, FID scores, or dataset numbers.

Key Findings:

Independently trained image generators converge to similar latent spatial organization, suggesting an implicit shared geometry
Intermediate denoising states can be swapped between unaligned models while preserving coarse visual content, even without explicit alignment
A single shared-space interface (with per-model LAEs) is sufficient to replace pairwise cross-model mappings across four distinct downstream tasks

Technical Novelty: Prior cross-model distillation and translation methods require pairwise mappings between every source/target model pair (O(N²) complexity). InstantFusion introduces model-specific Latent AutoEncoders (LAEs) projecting into a single shared latent space (O(N) complexity), constrained jointly by (a) noise-level-matched state alignment and (b) local denoising velocity synchronization—the velocity-matching objective for trajectory consistency appears to be the key novel ingredient.

What's New: Rather than aligning generators pairwise or distilling one into another, the paper identifies and exploits a pre-existing cross-model latent compatibility and formalizes it via a shared interface. The combined state-alignment + velocity-synchronization training objective targets both the geometry and the dynamics of denoising trajectories simultaneously, which prior cross-model work treats separately.

Extension Opportunities:

Extend the shared latent interface to video diffusion models, where cross-model handoffs could enable mixing motion priors from one model with appearance priors from another
Apply InstantFusion to multimodal generators (text-to-3D, text-to-audio) to test whether the shared latent hypothesis generalizes beyond 2D image generation
Use the LAE interface as a plug-in router for inference-time model selection, dynamically switching generators mid-trajectory based on content type (e.g., faces vs. landscapes)

Replicability: The abstract does not mention code or data release. Reproduction would require training LAEs for each target generator plus access to multiple frozen pretrained diffusion/flow models (likely SD, SDXL, Flux-class). Compute is likely moderate—LAEs are described as 'lightweight,' so training should be feasible on single-node multi-GPU setups rather than requiring full generator retraining.

Research Gaps:

Cross-model alignment of latent representations across heterogeneous diffusion/flow architectures was underexplored
Consistency of denoising dynamics (not just endpoints) across independently trained generators had not been systematically characterized or leveraged

🦾 ROBOTICS

1. ATHLETE: Learning Reactive Humanoid Tennis via Trajectory-Guided Motion Matching

Authors: Zhihao Tian, Zifan Wang, Enzheng Zhu... Published: 2026-10-09 | Citations: 0 arXiv | PDF

Research Question: How can humanoid robots learn reactive whole-body tennis skills without costly task-specific motion-capture data covering diverse approach states, contact locations, and timing conditions?

Summary: ATHLETE learns reactive humanoid tennis by combining generic locomotion data with ordinary tennis videos through Trajectory-Guided Motion Matching, avoiding costly task-specific mocap. A teacher-student pipeline with DAgger distillation and PPO fine-tuning yields a reactive policy achieving 90%/86.67% forehand/backhand hit rates on real hardware with multi-shot human rallies.

Key Results: ATHLETE achieved 90.00% forehand hit rate and 86.67% backhand hit rate on a real humanoid robot, and successfully sustained multi-shot rallies with human players. Training combined LAFAN1 locomotion dataset with video-derived striking motions.

Key Findings:

Generic locomotion data (LAFAN1) plus web tennis videos can substitute for expensive task-specific mocap when composed via TGMM
Teacher-student distillation with online DAgger produces a policy that operates purely from robot/ball observations at deployment
The approach transfers to real hardware with high hit rates and sustains multi-shot rallies with humans

Technical Novelty: Trajectory-Guided Motion Matching (TGMM) composes generic locomotion data (LAFAN1) with video-derived striking motions into task-conditioned whole-body references, then distills via teacher-student with online DAgger followed by PPO fine-tuning — enabling a deployment policy that needs no reference-motion input at runtime.

What's New: Unlike prior humanoid sports work that depends on dense task-specific demonstrations, ATHLETE synthesizes task-conditioned references by matching generic locomotion to video-derived striking trajectories, enabling reactive deployment without reference-motion inputs.

Extension Opportunities:

Extend TGMM to other racket/bat sports (table tennis, badminton, baseball) where similar reactive striking dynamics apply
Replace video-derived striking priors with large-scale internet video pretraining to eliminate curated demonstration requirements
Add opponent modeling and strategic shot selection on top of the reactive hitting policy to enable competitive rally play

Replicability: Code and result videos available at anonymous.4open.science/w/ATHLETE2027/. LAFAN1 is public; video-derived striking data likely custom. Compute not specified but PPO + DAgger on humanoid sim typically requires multi-GPU days.

Research Gaps:

No reported comparison against baselines using task-specific mocap or alternative motion-synthesis methods
Generalization to varied ball speeds, spin, opponent styles, and outdoor court conditions is not characterized

2. AutoBio+: Benchmarking Dexterous Manipulation with Dynamic Experimental Phenomena in Realistic Laboratories

Authors: Xuanye Wu, Yue Shi, Yongcan Yang... Published: 2026-10-09 | Citations: 0 arXiv | PDF

Research Question: How can we systematically benchmark robotic dexterous manipulation in realistic biological laboratory settings that include dynamic visual phenomena (flames, vapor, bubbles) and photorealistic environments?

Summary: AutoBio+ is a simulation benchmark for dexterous robotic manipulation in biological labs, featuring 14 tasks over 4 embodiments, controllable dynamic phenomena (flames, vapor, bubbles), and a hybrid Gaussian-splatting + simulated-foreground renderer. A Parallel Finite-State Machine on mjlab enables scalable demo generation and closed-loop policy evaluation, with real-world experiments indicating a reduced sim-to-real gap.

Key Results: Introduces AutoBio+ with 14 tasks across 4 robot embodiments, backgrounds reconstructed via Gaussian splatting from 6 real lab scenes, built on mjlab with a Parallel Finite-State Machine (PFSM) for batched rendering and decoupled trajectory replay. Evaluation shows substantial variance in policy performance across tasks, and real-world experiments demonstrate a reduced sim-to-real gap.

Key Findings:

Current manipulation policies show substantial per-task performance variance, exposing distinct capability gaps in lab settings
Integrating Gaussian-splatting backgrounds from 6 real labs with simulated interactive foregrounds narrows the sim-to-real gap
PFSM with decoupled trajectory replay and batched rendering enables efficient large-scale demonstration generation and model-independent evaluation

Technical Novelty: Hybrid rendering combining interactive simulated foreground with Gaussian-splatting reconstructed lab backgrounds, plus a Parallel Finite-State Machine (PFSM) on top of mjlab that decouples trajectory replay from batched rendering to enable scalable, model-independent closed-loop evaluation under controllable dynamic phenomena.

What's New: First benchmark to combine dexterous lab manipulation, controllable dynamic experimental phenomena, and Gaussian-splatting reconstructed real-lab backgrounds within a parallelized FSM-driven evaluation framework.

Extension Opportunities:

Add more dynamic phenomena (liquid pouring, color-changing reactions, precipitation) and physically-coupled chemistry simulators to broaden task realism
Train and benchmark VLA foundation models (e.g., OpenVLA, π0) on AutoBio+ to establish stronger baselines for lab-specific manipulation
Extend to multi-robot collaborative lab workflows (handoffs between arm, mobile base, humanoid) and long-horizon protocol execution with LLM task planners

Replicability: Supplementary materials and project page at autobio-plus.github.io/AutoBioPlus (code availability implied but not explicitly stated in abstract). Reproducing requires mjlab, GPU-based Gaussian splatting rendering pipeline, and 4 robot embodiment assets; large-scale parallel simulation suggests multi-GPU compute for demonstration generation.

Research Gaps:

Lack of baselines for state-of-the-art VLA/foundation policies and no reported absolute success-rate numbers in the abstract
Dynamic phenomena are visually controllable but physical coupling with task outcomes (e.g., chemical reactivity) remains limited

3. LAPT-Scale: Scaling Embodied Policies with Latent Action Pre-Training on Human Videos

Authors: Zhiyuan Xiang, Peidong Liu, Mingyang Li... Published: 2026-10-09 | Citations: 0 arXiv | PDF

Research Question: How can we scale generalist robotic policies without the prohibitive cost of collecting labeled embodied interaction data, by leveraging unlabeled egocentric human videos as a supervision source?

Summary: LAPT-Scale is a latent action pre-training paradigm that learns continuous latent actions from 23.3K hours of unlabeled egocentric human videos and uses them as implicit supervision for training vision-language-action (VLA) robotic policies. The approach demonstrates clear scaling behavior, lifting real-robot success from 63.3% to 85.0% as annotated video grows from 1K to 5K hours.

Key Results: Trained on 23.3K hours of egocentric video, LAPT-Scale improves real-robot success from 63.3% to 85.0% as annotated human video scales from 1K to 5K hours, demonstrating consistent scaling gains. The learned policy also generalizes to unseen visual conditions.

Key Findings:

Continuous latent action spaces outperform vector-quantized LAMs by preserving fine-grained manipulation dynamics
Combining perceptual, semantic, and optical-flow alignment signals improves transition-level supervision beyond pixel reconstruction
Policy performance scales predictably with human video data volume (63.3% → 85.0% success as data grows 1K→5K hours)
Human-video-pretrained policies generalize to unseen visual conditions on real robots

Technical Novelty: Unlike prior vector-quantized Latent Action Models (LAMs), LAPT-Scale uses a continuous latent space to preserve fine-grained motion dynamics, and combines perceptual + semantic + optical-flow alignment signals for transition supervision rather than relying solely on pixel reconstruction. The three-stage pipeline (human video LAM pre-training → embodiment alignment → task post-training) is a distinct scaling recipe.

What's New: First work to combine continuous (non-quantized) latent action modeling with multi-signal transition alignment at 23.3K-hour scale, and to demonstrate clean scaling laws for human-video-to-robot policy transfer.

Extension Opportunities:

Apply the continuous latent action model to bimanual or dexterous manipulation tasks requiring finer motion fidelity than vector-quantized LAMs can capture
Extend the perceptual/semantic/optical-flow alignment framework to third-person video sources (YouTube, Ego4D variants) to further scale pre-training data beyond egocentric collections
Investigate cross-embodiment transfer by using LAPT-Scale-LAM as a shared latent space for multiple robot morphologies, reducing per-robot alignment data requirements

Replicability: The abstract does not mention code or data release. Reproducing at full scale would require 23.3K hours of egocentric video (likely Ego4D/EPIC-Kitchens-scale datasets), substantial GPU compute for VLA pre-training (likely multi-node clusters), and real-robot hardware for the final evaluation stage — making full replication difficult outside well-resourced labs.

Research Gaps:

No comparison provided (in abstract) against recent VLA baselines like OpenVLA or RT-2 to contextualize the 85.0% success rate
Scaling curve stops at 5K hours despite 23.3K hours of training data — the ceiling of the scaling trend remains unclear

💻 COMPUTE

1. Deploying RDMA for Large-Scale Inference and Training Workloads

Authors: Yibo Huang, Yi Cui, Yukai Huang... Published: 2026-10-08 | Citations: 0 arXiv | PDF

Research Question: How can RDMA be effectively deployed in production parameter-server-based ML inference and training systems, given that hardware offload alone cannot handle variable-size messages, concurrent connections, bursty incasts, and service failures?

Summary: This paper presents ByteExpress, a production RDMA software communication layer deployed at ByteDance for parameter-server-based ML inference and training. It demonstrates that realizing RDMA's benefits in production requires software-layer solutions for buffer management, connection scalability, and failure recovery—not just hardware offload—yielding substantial latency and CPU reductions in real deployment.

Key Results: ByteExpress deployed at ByteDance demonstrated: (1) 35.7% reduction in mean model parameter fetch latency and 43.7% reduction in P99 latency for inference vs kernel TCP; (2) 10.69% reduction in total CPU usage; (3) up to 74.0% reduction in P99 model parameter push latency during training; (4) zero PFC (Priority Flow Control) pauses observed at runtime in production.

Key Findings:

Efficient RDMA requires coordinated buffer management across application and transport layers (not just hardware offload)
Scalable communication requires joint management of connections, receive capacity, and outstanding data via QP multiplexing and credit-based flow control
Service availability in production demands recovery mechanisms above the transport layer (RPC-assisted recovery) since hardware alone cannot handle service failures

Technical Novelty: Novel integrated software layer combining: (1) cross-layer RDMA-aware buffer allocation with zero-copy transfers spanning application and transport layers; (2) QP (Queue Pair) multiplexing with receive-credit flow control and software-level send windows for scalable connection management; (3) RPC-assisted recovery above the transport layer—a departure from pure hardware offload approaches that cannot handle production-scale service failures.

What's New: Unlike academic RDMA studies that focus on microbenchmarks or single components, this paper reports battle-tested production lessons from deploying RDMA at ByteDance scale. The integrated co-design of allocator, QP multiplexing, software send window, and RPC-assisted recovery—motivated by real failure modes—distinguishes it from prior point solutions.

Extension Opportunities:

Apply ByteExpress's RDMA-aware allocator and QP multiplexing techniques to all-reduce collective communication patterns used in modern LLM training (vs parameter-server architecture)
Extend the RPC-assisted recovery mechanism to handle multi-tenant GPU cluster failures where multiple training jobs share RDMA fabric
Integrate ByteExpress's software send window and receive-credit flow control into open-source frameworks like NCCL or Gloo to make these optimizations broadly accessible

Replicability: No code availability mentioned in the abstract. Reproduction would require RDMA-capable hardware (RoCE/InfiniBand NICs), a datacenter-scale parameter server deployment, and ML training/inference workloads at production scale—likely infeasible for academic reproduction without industry partnership.

Research Gaps:

No discussion of how these techniques generalize beyond parameter-server architectures to modern collective-communication-based training (e.g., all-reduce for LLMs)
Limited exploration of behavior under heterogeneous network conditions or multi-vendor RDMA NIC environments

2. GPU-resident Execution for Timing-predictable Sensor Processing Systems

Authors: Soham Sinha, Rahul Choudhury, Shekhar Dwivedi... Published: 2026-10-09 | Citations: 0 arXiv | PDF

Research Question: How can real-time sensor-processing pipelines eliminate host-mediated latency and jitter from CPU orchestration, I/O interrupts, and CUDA API overheads in GPU-accelerated closed-loop systems?

Summary: The paper introduces GPU-resident Execution, a model that keeps sensor I/O and processing inside a persistent GPU loop so the CPU never mediates each iteration. Implemented in NVIDIA Holoscan, it slashes closed-loop jitter by 7–170x on real camera-to-display workloads, making GPU-accelerated real-time systems far more timing-predictable.

Key Results: On an industry-grade 1080p/60Hz camera-to-display pipeline, GPU-resident Execution reduced photon-to-display jitter from 4.83ms to 0.66ms (7.3x improvement) versus a CPU-driven GPU-memory RDMA baseline, and cut software-processing jitter from 1.70ms to 0.01ms (170x). Evaluated across NVIDIA GPUs/SoCs on inference, host-contention, multi-loop, and streaming camera workloads using the NVIDIA Holoscan C++ framework.

Key Findings:

Host-mediated CUDA pipelines retain significant jitter even with RDMA and core isolation because the CPU scheduler and CUDA API remain in the critical path
A device-resident loop with GPU-visible I/O endpoints removes host wakeups entirely and achieves two-orders-of-magnitude reductions in software-processing jitter
The model is implementable as a drop-in set of components within an existing production framework (Holoscan) rather than requiring a new runtime

Technical Novelty: A long-lived device-resident execution loop that keeps I/O ingress, kernel processing, and output egress entirely on the GPU — removing host wakeups from the critical path. Prior host-side optimizations (GPU-memory RDMA, CPU core isolation) still relied on the CPU to submit CUDA work per iteration; this work makes the GPU itself the orchestrator.

What's New: First execution model to make the GPU itself the long-lived orchestrator of real-time sensor loops, superseding host-centric optimizations that still paid CPU-scheduling and CUDA-launch overheads each iteration.

Extension Opportunities:

Port the GPU-resident model to AMD ROCm or Intel oneAPI to generalize beyond NVIDIA's CUDA/Holoscan ecosystem
Extend to multi-GPU device-resident pipelines with GPU-to-GPU peer handoff for distributed robotic/autonomous workloads
Add formal timing-analysis / WCET bounds for GPU-resident kernels so safety-certification (ISO 26262, DO-178C) can rely on this execution model

Replicability: Implemented as reference components in NVIDIA's open-source Holoscan C++ SDK, which suggests code availability. Reproduction requires NVIDIA discrete GPUs and/or Jetson-class SoCs plus a camera/sensor source capable of GPU-direct DMA (e.g., Rivermax or GPUDirect RDMA NIC).

Research Gaps:

No formal worst-case timing guarantees — strong empirical jitter reductions but no analytical bounds suitable for safety certification
Vendor lock-in: approach is demonstrated only on NVIDIA stacks (CUDA + Holoscan); portability to other accelerators and heterogeneous SoCs is unaddressed

3. Minions: A Dynamic and Distributed Virtual GPU Cloud Built on Idle Apple Silicon Devices

Authors: Jianhua Li, Yun Xing, Girisha De Silva... Published: 2026-10-09 | Citations: 0 arXiv | PDF

Research Question: How can organizations overcome the cost, elasticity, and wait-time limitations of centralized GPU infrastructure for ML inference by harnessing idle GPU capacity in existing Apple Silicon macOS fleets?

Summary: Minions is a distributed virtual GPU cloud that pools idle Apple Silicon GPUs across macOS devices in an organization to serve large-scale ML inference. It provides dynamic scheduling, fault tolerance under device churn, and efficient utilization without requiring new GPU hardware procurement.

Key Results: The abstract reports qualitative evidence rather than specific numbers: the authors demonstrate that idle macOS devices can be aggregated into a functioning distributed GPU inference platform providing flexible task execution, fault tolerance, and efficient resource utilization. No concrete benchmarks, throughput figures, latency measurements, device counts, or model-specific results are disclosed in the abstract.

Key Findings:

Idle Apple Silicon GPUs in typical macOS fleets represent a substantial untapped inference resource.
A dynamic distributed scheduler can tolerate device churn (users reclaiming machines) while sustaining inference workloads.
Aggregated consumer-class GPUs can serve as a practical alternative when centralized datacenter GPU access is cost- or capacity-constrained.

Technical Novelty: Prior volunteer-compute systems (BOINC, Folding@home, Petals) targeted CPUs or NVIDIA GPUs in homogeneous server contexts. Minions is novel in specifically targeting the unified-memory Apple Silicon GPU (via Metal/MLX) as a first-class inference substrate and treating a corporate macOS fleet as an elastic virtual GPU cloud with dynamic scheduling and fault tolerance for churn from user activity reclaiming the device.

What's New: First system (as described) to treat Apple Silicon unified-memory GPUs across a macOS fleet as an elastic, fault-tolerant virtual GPU cloud for ML inference, rather than relying on commodity NVIDIA volunteer compute or centralized clusters.

Extension Opportunities:

Extend the aggregation layer beyond Apple Silicon to a heterogeneous mix (CUDA, ROCm, Intel Arc, mobile NPUs) with a unified MLX/MPS/CUDA abstraction and scheduler-aware kernel routing.
Add training and fine-tuning support (e.g., federated LoRA or pipeline-parallel QLoRA) on top of the inference-only substrate, with gradient compression tuned for consumer-grade intermittent networks.
Build a market/incentive layer (token-metered compute credits or privacy-preserving attestation) so idle devices across organizational boundaries can contribute, turning Minions into a decentralized inference marketplace.

Replicability: The abstract does not mention open-source code, released datasets, or a public artifact. Reproduction would require a fleet of Apple Silicon Macs (M1/M2/M3+ class), a coordinator service, and macOS-side agents using Metal/MLX; realistic evaluation needs tens of devices and representative LLM/vision workloads.

Research Gaps:

Abstract omits quantitative benchmarks — throughput, latency, cost-per-token, and comparison baselines against centralized A100/H100 or MLX-local inference are not reported.
No discussion of security, data confidentiality, or trust model for running sensitive inference payloads on employee workstations.

⚡ ENERGY

1. Doubly Metastable Synthesis for Accelerated Quantum Materials Discovery

Authors: Kyuho Lee, Berit H. Goodge, Soonhyun Kwon... Published: 2026-10-09 | Citations: 0 arXiv | PDF

Research Question: How can researchers expand the metastable synthesis window to access novel quantum materials phases that are inaccessible via conventional single-metastable synthesis routes?

Summary: The paper introduces diffusion-activated epitaxy, a synthesis technique that stacks two metastable methods (diffusion and epitaxy) to enlarge the accessible metastable phase space for quantum materials. The authors demonstrate the approach by stabilizing four phases in a narrow synthesis window, three of which have crystal structures absent from all known experimental and computational databases.

Key Results: Within a narrow window of investigated synthesis space, the authors stabilized four distinct phases using diffusion-activated epitaxy. Three of these four phases exhibit XRD and STEM characterization signatures incompatible with any crystal structures registered in experimental or computational inorganic structure databases, demonstrating access to previously unrealized phase space.

Key Findings:

Combining diffusion and epitaxy compounds their individual metastable windows, enabling access to phases inaccessible to either method alone
Four distinct phases were stabilized within a narrow synthesis parameter range, indicating a rich accessible phase landscape
Three of the four phases are structurally novel relative to all experimental and computational inorganic structure databases

Technical Novelty: The combination of two independent metastable synthesis handles — solid-state diffusion and epitaxial strain/templating — into a single 'doubly metastable' technique (diffusion-activated epitaxy), compounding the kinetic stabilization windows rather than relying on either alone.

What's New: Prior metastable synthesis work typically exploits a single kinetic handle (either diffusion-limited reactions or epitaxial templating). This paper explicitly compounds two orthogonal metastable mechanisms, framing 'double metastability' as a general design principle for materials discovery rather than a one-off trick.

Extension Opportunities:

Apply diffusion-activated epitaxy to other quantum materials families (e.g., unconventional superconductors, topological materials, or 2D magnets) to systematically map the enlarged metastable window
Build a high-throughput computational screening pipeline that predicts which precursor/substrate combinations maximize the doubly-metastable window, feeding candidates into automated synthesis
Solve the crystal structures of the three novel database-incompatible phases and characterize their electronic/magnetic properties to identify new physics

Replicability: The abstract does not mention open code or data releases. Reproduction would require a thin-film epitaxy setup (MBE or PLD), high-temperature diffusion capability, synchrotron/lab XRD, and aberration-corrected STEM — a well-equipped condensed-matter materials lab, likely $5–20M in capital equipment.

Research Gaps:

The abstract does not disclose the specific material system or the physical properties (electronic, magnetic, superconducting) of the novel phases
No systematic theoretical framework yet predicts which material systems will benefit most from doubly metastable synthesis or how large the enlarged window can grow

2. Magnetic-field-induced detwinning of orientational domains in altermagnet $α$-MnTe: A single-crystal neutron diffraction study

Authors: Feihao Pan, Yishui Zhou, Sabreen Hammouda... Published: 2026-10-09 | Citations: 0 arXiv | PDF

Research Question: How can one prepare a well-defined, near-single-orientation magnetic domain state in bulk altermagnet α-MnTe to overcome domain averaging that obscures intrinsic chiral magnon and spin-split responses?

Summary: The authors use single-crystal neutron diffraction to show that an in-plane magnetic field along [-1,1,0] detwins the three orientational magnetic domains of hexagonal altermagnet α-MnTe, saturating near 2 T. A constrained refinement at 2 K/2 T gives a nearly single-orientation state (97% population) with M ∥ [1,1,0] and an ordered moment of 4.62(8) μ_B/Mn, providing a non-contact route to prepare well-defined bulk altermagnetic states.

Key Results: Single-crystal neutron diffraction under an in-plane field along [-1,1,0] showed domain alignment saturating near μ₀H = 2 T. Constrained refinement of 61 nuclear and magnetic reflections at 2 K and 2 T yielded M ∥ [1,1,0], an ordered moment of 4.62(8) μ_B per Mn, and orientational domain populations r_A:r_B:r_C = 0.97:0.03(5):0. Anomalous behavior in reflection intensities was observed between 60–88 K.

Key Findings:

In-plane field along [-1,1,0] produces a dominant orientational domain population of 0.97:0.03(5):0, saturating near μ₀H = 2 T
Refined Néel-vector direction is M ∥ [1,1,0] with ordered Mn moment of 4.62(8) μ_B at 2 K
Anomalous temperature dependence of selected reflections between 60–88 K suggests possible intrinsic piezomagnetic or magnetoelastic effects
Unpolarized diffraction cannot distinguish time-reversal antiphase domains, so the state is near-single-orientation rather than truly single-domain

Technical Novelty: First demonstration of non-contact in-plane magnetic field detwinning of orientational domains in bulk α-MnTe quantified by single-crystal neutron diffraction refinement, complementing existing strain and field-cooling approaches.

What's New: Prior work on α-MnTe relied on strain or field-cooling to control domains; this paper quantifies direct in-plane field-induced detwinning via single-crystal neutron refinement and simultaneously flags a previously underreported 60–88 K reflection anomaly attributable to magnetoelastic/piezomagnetic coupling.

Extension Opportunities:

Use polarized neutron diffraction to resolve time-reversal antiphase (180°) domains that unpolarized diffraction cannot distinguish, achieving a truly single-domain state
Combine field-induced detwinning with chiral magnon spectroscopy (inelastic neutron scattering or THz probes) to measure the predicted non-reciprocal magnon dispersion in a near-single-domain α-MnTe sample
Systematically map the 60–88 K anomaly via combined magnetostriction/piezomagnetic measurements under field to isolate magnetoelastic coupling constants and test for an intrinsic piezomagnetic effect

Replicability: No code/data link mentioned in the abstract. Reproduction requires access to a single-crystal neutron diffractometer with a horizontal-field cryomagnet (≥2 T, base T ~2 K) and high-quality α-MnTe single crystals — specialized large-facility (reactor/spallation) resources.

Research Gaps:

Unpolarized neutron diffraction cannot distinguish domain repopulation from continuous Néel-vector rotation, nor resolve time-reversal (180°) antiphase domains
The microscopic origin of the 60–88 K anomaly — intrinsic piezomagnetic vs. extrinsic magnetoelastic — remains unresolved

3. Giant geometric nonlinear transport in inhomogeneously strained Bi$_2$Se$_3$ single crystalline slabs

Authors: Igor Veremchuk, Pavlo Makushko, Anita Guarino... Published: 2026-10-09 | Citations: 0 arXiv | PDF

Research Question: How can nonlinear electrical responsivity in non-centrosymmetric materials be dramatically enhanced beyond the limits of thin-film topological insulators to enable practical zero-threshold energy harvesters and frequency upconverters?

Summary: The paper reports a six-orders-of-magnitude boost in nonlinear electrical responsivity (reaching 10^4 μm/V) in macroscopic Bi2Se3 topological-insulator slabs subjected to inhomogeneous strain, attributing the giant effect to a classical geometric nonlinear transport mechanism driven by bulk strain gradients. It generalizes a phenomenon previously seen only in microscale-patterned structures to bulk crystals, establishing macroscopic strain engineering as a design principle for nonlinear transport devices.

Key Results: Demonstrated a record-high nonlinear responsivity of 10^4 μm/V in inhomogeneously strained macroscopic Bi2Se3 single-crystalline slabs — a six-orders-of-magnitude improvement over Bi2Se3 thin-film counterparts. Attributed to a classical geometric nonlinear transport effect driven by complex bulk strain gradients, extending a mechanism previously confined to microscale-patterned structures to macroscopic samples.

Key Findings:

Inhomogeneously strained macroscopic Bi2Se3 slabs reach a record nonlinear responsivity of 10^4 μm/V, exceeding thin-film values by ~10^6
The enhancement is explained by a classical geometric nonlinear transport effect arising from complex bulk strain gradients, not an intrinsic Berry-curvature-dipole mechanism
The geometric nonlinear Hall mechanism, previously demonstrated only in microscale-patterned devices, operates in macroscopic bulk samples under strain inhomogeneity

Technical Novelty: Prior geometric nonlinear Hall effects required microscale-patterned device geometries to break symmetry locally; this work shows that bulk inhomogeneous strain gradients in macroscopic crystals produce the same geometric mechanism, unlocking a six-orders-of-magnitude responsivity gain over thin films.

What's New: First demonstration that bulk strain-gradient inhomogeneity in a macroscopic topological-insulator crystal can realize the geometric nonlinear Hall mechanism and produce record responsivity, replacing the need for microscale device patterning with scalable strain engineering.

Extension Opportunities:

Apply the same macroscopic strain-gradient engineering to transition metal dichalcogenides (e.g., WTe2, MoTe2) and quantify responsivity scaling vs. strain gradient magnitude
Prototype a zero-threshold RF/microwave energy harvester or broadband frequency upconverter using strained Bi2Se3 slabs and benchmark conversion efficiency against Schottky-diode rectifiers
Engineer controlled strain-gradient patterns (e.g., via lithographically defined substrates or bending jigs) in moiré heterostructures to tune and spatially map the geometric nonlinear Hall response

Replicability: No code/data availability is stated in the abstract. Reproduction would require Bi2Se3 single-crystal growth (Bridgman or vapor transport), mechanical strain application with gradient control, and standard second-harmonic lock-in transport measurement — moderate materials-lab infrastructure, no significant compute.

Research Gaps:

No quantitative mapping of strain-gradient profile to local nonlinear response (strain tomography) is reported
Generality across other layered systems (TMDs, moiré stacks) is proposed but not experimentally demonstrated

🏥 HEALTHCARE

1. Designing Protein Structures from Natural Language

Authors: Joseph Lee, Yanjing Li, Bo Qiang... Published: 2026-10-09 | Citations: 0 arXiv | PDF

Research Question: How can we build a protein structure generation model that supports broad, flexible design objectives specified in natural language, rather than requiring a dedicated input representation for each property?

Summary: Prompteina is the first all-atom protein structure generation model conditioned on natural language, trained on a new 1.8M-structure, 50M-caption paired dataset by adapting the Proteina-Complexa diffusion backbone. It outperforms baselines on structural property control and binder design, and is experimentally validated with 9 novel binders against GM2A (strongest K_D ≈ 107 nM), establishing text as a general-purpose interface for de novo protein design.

Key Results: Prompteina was trained on Prompteina-Data (1.8M structures, 50M+ captions) by adapting the pretrained Proteina-Complexa model. It outperforms baselines on 15 of 18 structural properties, leads baselines on at least 13 of 20 binder targets, and improves interface confidence by 14% over the strongest baseline. With iterative in silico feedback, its gain over non-iterative generation is 57% larger than the next-best baseline on the hardest targets. Experimental validation against GM2A (a target with no reported de novo binders) yielded 9 binders, strongest at apparent K_D of 107 nM.

Key Findings:

Natural language can serve as a unified conditioning interface replacing per-property bespoke inputs for protein design
Prompteina beats baselines on 15/18 structural properties and ≥13/20 binder targets, with +14% interface confidence over the strongest baseline
Iterative in-silico feedback amplifies Prompteina's advantage by 57% over the next-best baseline on hard targets, and 9 experimentally validated GM2A binders (K_D down to 107 nM) confirm real-world utility

Technical Novelty: First all-atom structure generation model conditioned on free-form natural language at both global and per-residue granularity. Novel contributions include: (1) the Prompteina-Data paired text-structure corpus with residue-aligned captions, and (2) a two-stage adaptation (text-conditioned pretraining + residue-text alignment finetuning) applied to a frozen structure diffusion backbone (Proteina-Complexa).

What's New: Prior protein design models require dedicated input channels for each property (motif scaffolding, symmetry, etc.). Prompteina is the first to treat protein design analogously to text-to-image generation, with a single text interface controlling both global and residue-level properties — supported by a purpose-built paired text-structure dataset at unprecedented scale.

Extension Opportunities:

Extend the caption vocabulary to cover dynamic/conformational properties (allostery, folding intermediates) rather than static structural/functional descriptors
Integrate active-learning loops where wet-lab binding assay results feed back into the text-conditioned finetuning to iteratively refine language-to-structure alignment
Combine Prompteina with inverse-folding and codon-optimization pipelines to produce an end-to-end natural-language-to-expressible-protein design stack

Replicability: The abstract does not state code/data availability. Reproducing would require substantial compute: generating 50M+ captions over 1.8M structures (likely LLM-based captioning), plus adapting a large pretrained all-atom structure diffusion model — plausibly multi-node GPU training for weeks. Wet-lab validation (GM2A binder screening) requires biochemistry resources beyond the compute budget.

Research Gaps:

Caption quality and coverage bound the model: properties not well-described in training text (e.g., dynamics, catalytic mechanism details) likely remain inaccessible
Experimental validation is demonstrated on one target (GM2A); generalization of hit rates and affinities across diverse functional classes remains unproven

2. TopoLIA: Topology-Inspired Latent Interaction Attention for Drug Discovery

Authors: Xiang Liu, Yiming Ren, Mustafa Hajij... Published: 2026-10-09 | Citations: 0 arXiv | PDF

Research Question: How can transformer attention mechanisms be redesigned to better model complex heterogeneous interactions in drug discovery, where conventional source-target attention becomes a bottleneck for capturing multiscale structural and geometric patterns in protein-ligand complexes?

Summary: TopoLIA introduces a topology-inspired attention mechanism (LIA) that routes source-target information through a compact latent interaction space rather than direct attention, combined with persistent spectral tokens that encode multiscale geometry of 3D protein-ligand complexes. Pretrained via masked spectral token reconstruction, it achieves SOTA on protein-ligand scoring and docking benchmarks while demonstrating that compact latent dimensions act as an effective inductive bias for biomolecular learning.

Key Results: TopoLIA achieves state-of-the-art performance across multiple scoring and docking benchmarks for protein-ligand binding, consistently outperforming its standard-attention counterpart. Rank analyses demonstrate that compact latent dimensions yield the best performance. The model is pretrained via masked spectral token reconstruction on unlabeled molecular structures, yielding transferable representations. Specific numerical results (e.g., exact RMSE/Pearson correlation on PDBbind, CASF-2016 scoring/docking power) are not quoted in the abstract.

Key Findings:

Mediating attention through a compact latent interaction space outperforms standard source-target attention on protein-ligand benchmarks
Persistent topological spectra effectively capture multiscale geometric patterns in 3D biomolecular complexes as transformer input tokens
Compact (low-rank) latent dimensions yield best performance, suggesting compact interaction spaces are a useful inductive bias
Masked spectral token pretraining produces transferable structural representations without task-specific labels

Technical Novelty: The core novelty is latent interaction attention (LIA): instead of standard source-target attention computing direct Q-K-V products, information is routed through a compact learned latent interaction space that governs both how much and how information is transformed. This is combined with topological spectral tokens (persistent spectra) as the input representation, and a masked spectral token reconstruction pretraining objective — a departure from standard atom-type or 3D-coordinate masking.

What's New: Unlike prior transformer approaches that use direct source-target attention or geometric/graph-based message passing, TopoLIA fuses two ideas: (1) a latent-mediated attention that structurally constrains interaction representations, and (2) topological spectral features (persistent spectra) as native tokens — integrating topology into both representation and architecture rather than as a post-hoc feature.

Extension Opportunities:

Apply the LIA framework beyond protein-ligand binding to protein-protein interactions, antibody-antigen docking, or RNA-small molecule binding where heterogeneous interaction modeling matters
Replace persistent spectra with other topological descriptors (persistent Laplacians, persistent sheaves, or Mapper-based features) and benchmark whether the inductive bias generalizes
Extend the latent interaction attention to multi-modal scientific problems like materials science, climate modeling, or multi-agent reinforcement learning where source-target information exchange is central

Replicability: The abstract does not mention a code or data release. Reproduction would require PDBbind-scale protein-ligand datasets, persistent spectra computation (likely via GUDHI or similar TDA libraries), and transformer pretraining compute — probably multi-GPU for days given masked-reconstruction pretraining on unlabeled structures. Moderate-to-high compute barrier; code availability unconfirmed.

Research Gaps:

No reported comparison to recent 3D-equivariant architectures (e.g., EquiBind, DiffDock, Uni-Mol) or quantitative numbers in the abstract to contextualize SOTA claims
Unclear how LIA scales to very large interaction partners (e.g., full protein complexes) or whether the compact latent space becomes a bottleneck for highly heterogeneous systems

3. RetroTAC: Learning Route-Derived Synthetic Accessibility for PROTACs

Authors: Stefano Ribes, Jia Xin, Zhu Zhu... Published: 2026-10-09 | Citations: 0 arXiv | PDF

Research Question: How can we rapidly estimate the synthetic accessibility of PROTACs (proteolysis-targeting chimeras) without running expensive full retrosynthesis planning, while avoiding the over-penalization of molecular complexity that traditional SA scores apply to these large modular molecules?

Summary: RetroTAC introduces a PROTAC-specific synthetic accessibility score that is derived from retrosynthetic route properties (depth, fragment-join balance) rather than generic complexity penalties, and a fast ML ensemble that predicts this score from structure alone. It achieves R²=0.65 against the explicit score and strongly discriminates synthesizable from unsynthesizable PROTAC candidates in virtual library enumeration (98.75% vs 51.00% fully resolved routes).

Key Results: The authors introduce a route-derived synthesizability score and RetroTAC, an ML ensemble surrogate predicting it directly from structure. RetroTAC achieves RMSE of 0.13 and R² of 0.65 on held-out molecules. In a virtual PROTAC library enumeration (recombining warheads, linkers, E3-ligase ligands), high-score selected candidates achieved a mean route-derived synthesizability of 0.73 vs 0.31 for low-score candidates, with 98.75% vs 51.00% fully resolved under tested retrosynthetic search settings.

Key Findings:

A route-derived score rewarding short depth and balanced fragment-joining captures PROTAC synthesizability better than complexity-penalizing SA scores
The ensemble surrogate predicts the score with RMSE 0.13 and R² 0.65, enabling fast library-scale prioritization
On enumerated PROTACs, high-predicted-score candidates were nearly fully resolvable by retrosynthesis (98.75%) versus only 51.00% for low-score candidates, validating practical screening utility

Technical Novelty: Prior SA scores (SAScore, SCScore) penalize molecular complexity universally and overestimate PROTAC difficulty, while retrosynthesis planners (AiZynthFinder, Chemformer) are accurate but too slow for library-scale screening. The novelty is twofold: (1) a PROTAC-tailored interpretable scoring function that rewards short synthetic depth and balanced fragment joining — exploiting PROTACs' modular warhead-linker-E3 architecture — and (2) an ensemble ML surrogate (RetroTAC) that predicts this route-derived score directly from structure, bypassing the planning step.

What's New: First PROTAC-oriented synthesizability model that unifies interpretability (explicit route-depth and fragment-balance features) with ML speed, and explicitly targets the modular warhead-linker-E3 architecture rather than treating PROTACs as generic complex molecules.

Extension Opportunities:

Integrate RetroTAC as a scoring filter inside PROTAC generative models (e.g., diffusion/VAE linker generators) to steer sampling toward synthesizable candidates during generation rather than only post-hoc filtering
Extend the route-derived score formulation to other modular chemistries like molecular glues, bifunctional antibody-drug conjugates, or peptide-small-molecule hybrids where modular assembly also dominates synthesis
Combine RetroTAC with wet-lab success-rate data to calibrate the score against actual synthesis outcomes, converting predicted depth/balance into probabilistic yield estimates

Replicability: The abstract does not explicitly mention public code or data release. Reproduction would require: a retrosynthesis planner (likely AiZynthFinder or similar) to generate training labels, a PROTAC enumeration dataset built from warhead/linker/E3-ligand libraries, and modest GPU compute to train ensemble models — tractable for a small research group but label-generation (route planning on thousands of molecules) is the main compute bottleneck.

Research Gaps:

No reported wet-lab validation — the score is validated against retrosynthesis planner output, not actual synthesis success rates or yields
Generalization to novel linker chemistries or beyond-rule-of-five PROTAC chemotypes not characterized; the enumeration-based evaluation may bias toward well-represented building blocks

🔬 MATERIALS

1. Electron- and hole-doped superconductivity in two-dimensional electride ZrX$_2$ (X = Cl, Br, I)

Authors: Tian-Xin Liu, Xin-Yang Li, Zuo-Wei Yin... Published: 2026-10-09 | Citations: 0 arXiv | PDF

Research Question: Can monolayer ZrX₂ (X = Cl, Br, I) 2D semiconducting electrides host phonon-mediated superconductivity under both electron and hole doping, and what mechanisms control the asymmetry between bipolar superconducting responses?

Summary: The paper uses first-principles calculations to predict phonon-mediated bipolar superconductivity in monolayer ZrX₂ (X = Cl, Br, I) electrides, finding electron doping produces high Tc up to 30 K at extremely low carrier concentrations (0.01 e/f.u.) driven by a flat conduction band minimum and nearby Van Hove singularities, while hole doping yields only weak superconductivity (~2 K) near a magnetic instability. Comparison with InSe and LaBr₂ establishes the spatial and orbital character of excess interstitial electrons as the key factor controlling the hole-doped response.

Key Results: First-principles calculations demonstrate phonon-mediated superconductivity in monolayer ZrX₂ under bipolar doping. Electron doping yields high Tc at ultra-low concentrations: ZrBr₂ Tc = 30.0 K at 0.01 e/f.u., ZrCl₂ Tc = 22.5 K at 0.05 e/f.u., ZrI₂ Tc = 18.1 K at 0.01 e/f.u. Hole doping requires 0.2–0.25 h/f.u. to activate superconductivity, yielding only Tc ~ 2 K, and approaches a magnetic instability at higher hole doping. Comparative analysis with InSe and LaBr₂ identifies excess-electron spatial location and orbital character as the key determinants.

Key Findings:

ZrBr₂ reaches Tc = 30.0 K at just 0.01 e/f.u. electron doping — a remarkably low threshold enabled by flat-band physics
Strong electron–hole asymmetry: hole doping requires 20× higher carrier density to achieve only ~2 K Tc, and competes with magnetic ordering
Spatial location and orbital character of the electride's excess electrons govern the strength of hole-doped superconductivity and competing orders, as revealed by InSe/LaBr₂ comparison

Technical Novelty: Identification of bipolar (electron and hole) phonon-mediated superconductivity in a 2D electride, with mechanistic framework linking the asymmetry to interstitial anion-electron spatial character — a dimension beyond conventional band-filling analysis. The flat conduction band minimum with nearby Van Hove singularities enables record-low doping thresholds (0.01 e/f.u.) for high Tc.

What's New: First demonstration of bipolar (electron- and hole-doped) phonon-mediated superconductivity in a 2D electride, combined with a predictive mechanistic framework tying superconducting asymmetry to the real-space anion-electron topology — moving beyond generic band-structure arguments for flat-band superconductors.

Extension Opportunities:

Experimentally synthesize monolayer ZrBr₂ via exfoliation or CVD and test the predicted 30 K Tc using ionic liquid gating for low-concentration electron doping
Extend the flat-band + Van Hove singularity design principle to other 2D electride families (Y₂C, Ca₂N derivatives) to search for even higher Tc
Investigate heterostructure stacking of ZrX₂ with magnetic 2D materials to explore the predicted magnetic instability regime for coexisting superconductivity and magnetism

Replicability: No code/data release mentioned in the abstract. Reproduction requires standard DFT/DFPT stack (VASP or QE + EPW/Wannier90) with Migdal–Eliashberg solvers for electron-phonon coupling. Compute: ~thousands of CPU-hours on an HPC cluster for full q-grid phonon and EPC convergence across three compositions and doping ranges.

Research Gaps:

No experimental validation yet; synthesis and gating of monolayer ZrX₂ remain open challenges
The nature of the magnetic instability at higher hole doping (ordering vector, order parameter, interplay with SC) is identified but not characterized

2. Anomaly induced transport in holographic superconductors

Authors: Ashis Tamang, Nishal Rai, Javier Esteban... Published: 2026-10-09 | Citations: 0 arXiv | PDF

Research Question: Does anomaly-induced transport (e.g., chiral magnetic effect) in a holographic superconductor remain confined to the anomalous current sector, or can spontaneous symmetry breaking (SSB) propagate anomaly effects into non-anomalous U(1) currents?

Summary: The paper studies a holographic superconductor with vector, axial, and an extra non-anomalous U(1) gauge field plus a condensing charged scalar, computing chiral magnetic conductivities via Kubo formulae in the probe limit. It shows that spontaneous symmetry breaking channels anomaly-induced transport into the non-anomalous current sector, and that this behavior is robust across both Δ=1 and Δ=3 scalar quantizations, matching analytic limits.

Key Results: Using a probe-limit holographic model with vector + axial gauge fields plus an extra non-anomalous U(1) and a charged scalar, the authors numerically construct superconducting backgrounds in both Δ=1 (O₁) and Δ=3 (O₂) quantizations. Applying Kubo formulae, they compute chiral magnetic conductivities and show the scalar condensate induces anomaly-driven response in the non-anomalous sector. Numerical results match analytical limits as O_i → 0 and O_i → ∞ in both theories, confirming the effect is quantization-independent.

Key Findings:

Spontaneous symmetry breaking causes anomaly-induced transport to appear in the non-anomalous U(1) sector, not just the anomalous currents
The leakage effect is independent of scalar quantization — it appears in both the O₁ (Δ=1) and O₂ (Δ=3) theories
Numerical chiral magnetic conductivities interpolate consistently between analytical O_i → 0 and O_i → ∞ limits in both theories

Technical Novelty: Prior holographic CME studies worked in normal phases or restricted anomaly transport to the anomalous current. This paper couples vector + axial + an additional non-anomalous U(1) with a condensing scalar, and demonstrates — via Kubo analysis in both standard and alternative scalar quantizations — that SSB leaks anomaly-induced transport into the non-anomalous sector.

What's New: First demonstration that SSB in a holographic superconductor extends anomaly-induced transport into a non-anomalous sector, and that this cross-sector effect is quantization-independent — a feature not identified in earlier holographic CME or superconductor studies.

Extension Opportunities:

Move beyond the probe limit to include gravitational backreaction and study anomaly-induced transport near the critical temperature with full metric dynamics
Extend to finite temperature/chemical-potential phase diagrams relevant to QCD matter or Weyl semimetals, mapping how the induced non-anomalous CME scales with condensate strength
Add momentum relaxation or a lattice to compute DC conductivities and compare predicted anomaly-leakage signatures against transport data in chiral superconductors

Replicability: No code or data release is mentioned in the abstract. Reproduction requires standard holographic numerics: solving coupled ODEs for the probe gauge/scalar system on an AdS background and evaluating linear-response Kubo integrals — tractable on a laptop with Mathematica/Python (SciPy BVP solvers).

Research Gaps:

Analysis is restricted to the probe limit, ignoring backreaction effects that could qualitatively modify transport near criticality
No connection yet to experimental systems (Weyl/Dirac superconductors, QGP) or predictions of measurable transport signatures

3. Dislocation mediated twin boundary faceting in $β$-Sn

Authors: John Israel, Ian Chesser, Khanh Dang... Published: 2026-10-09 | Citations: 0 arXiv | PDF

Research Question: How do twin boundaries (specifically (301) TBs) in β-Sn facet via interactions with dislocations, and can the resulting facet geometry be quantitatively predicted by elasticity theory?

Summary: The authors use MLIP-driven molecular dynamics to show that (301) twin boundaries in β-Sn facet by absorbing ⟨c⟩ dislocations, with facet length growing proportionally to the number of absorbed dislocations. They demonstrate that quantitative agreement with isotropic elasticity theory requires treating the asymmetric twin facet as having disclination (not just dislocation) character, and argue this absorption-driven growth explains the prevalence of (301) twins in deformed β-Sn.

Key Results: Using MD simulations with a recently developed machine-learned interatomic potential (MLIP), the authors demonstrate a ⟨c⟩ dislocation-absorption mediated faceting reaction along (301) TBs in β-Sn where facet length scales with the number of absorbed dislocations. Predicted facet length matches isotropic elasticity theory only when the disclination character of the asymmetric twin facet is included in the model.

Key Findings:

A ⟨c⟩ dislocation-absorption mechanism drives faceting along (301) twin boundaries in β-Sn, with facet length scaling linearly with the number of absorbed dislocations
Isotropic elasticity theory predicts facet length accurately only when the disclination character of the asymmetric twin facet is included
The ubiquity of (301) twins in β-Sn likely arises from their ability to grow rapidly through dislocation forests via this absorption-mediated faceting

Technical Novelty: First MD demonstration of a ⟨c⟩ dislocation-absorption-mediated faceting reaction on (301) TBs in β-Sn, and the explicit incorporation of disclination character (not just dislocation content) of asymmetric twin facets into isotropic elasticity predictions of facet length — reconciling atomistic simulation with continuum theory for a low-symmetry metal.

What's New: Combines a new MLIP for β-Sn with disclination-aware continuum analysis to mechanistically explain experimentally observed deformation-induced twin faceting in a low-symmetry (tetragonal) metal, bridging atomistic simulation and defect theory in a system where prior work was largely phenomenological.

Extension Opportunities:

Extend the MLIP-driven MD framework to multi-component Sn-based solder alloys (e.g., Sn-Ag-Cu) to study how alloying elements modify twin-dislocation absorption and facet growth kinetics
Couple the atomistic faceting results with mesoscale crystal plasticity or phase-field models to predict macroscopic strengthening and fatigue response of solder joints under thermomechanical cycling
Experimentally validate the disclination-corrected facet-length prediction using in-situ TEM nanoindentation on β-Sn single crystals and compare against the MD-predicted scaling law

Replicability: The abstract references a recently developed MLIP but does not state code/data availability. Reproduction would require the MLIP parameters, a standard MD engine (e.g., LAMMPS), and modest HPC resources (tens of thousands of core-hours typical for MLIP-driven defect-interaction simulations on multi-million-atom twin-dislocation systems).

Research Gaps:

No treatment of temperature, strain-rate, or alloying effects relevant to real solder service conditions
Analogies to faceting in other low-symmetry metals are hypothesized but not systematically quantified or generalized into a predictive framework

🔥 GitHub Trending

1. qwdwqfwq/topconf-paper-figure-gallery

⭐ 924 stars | HTML

3,516 curated Figure 1 / teasers from ICLR, ICML, NeurIPS, CVPR, ACL, AAAI (2023-2026), with tier badges and FigureForge retrieval-augmented figure drafting

aaai academic-writing acl awesome-list computer-vision cvpr

2. wfzyx/von

⭐ 889 stars | Python

The open-source System One decision model. Sub-15ms, non-autoregressive, local drop-in alternative to TypeSafe Jev.

decision-model jev machine-learning python rlcd system-one

3. whitecircle/halo

⭐ 821 stars | Python

Halo is an open-source framework built by White Circle for training large language and multimodal models

agent async-rl distributed-training fine-tuning grpo large-language-models

4. Calix-L/DanKS

⭐ 717 stars | Python

RL‑Empowered Small‑Scale Competitive Guandan Agent

card-games game-ai guandan pytorch reinforcement-learning

5. strands-labs/strands-decider

⭐ 591 stars | Python

A small, fast decision model, or system one model, for agentic workflows. Pick between options or rate on a scale faster than an LLM, with a calibrated confidence on every decision.

ai machine-learning natural-language-understanding python strands-agents strands-labs

6. AbdelStark/awesome-typesafe-jev

⭐ 582 stars | HTML

Awesome Jev: a source-backed field guide to TypeSafe's System One model, with SDKs, live demos, agent tools, and independent evaluations.

agent-skills ai-agents awesome awesome-jev awesome-list awesome-lists

7. Sidiora-Labs/Paxeer-X-Network

⭐ 571 stars | Rust

Paxeer X is a Distributed HyperState Machine for payments, code execution and intent routing. Designed for Machines and the Users operating them.

agent agent-client-protocol ai ai-agents blockchain machine-learning

8. timoncool/YuE2-Studio

⭐ 556 stars | TypeScript

Local AI song generator with an editable score — YuE2 on your GPU: full songs with vocals, sheet music, covers, exact replay. Native Windows app, no Python, installer with auto-update.

abc-notation ai-music covers cuda ggml local-ai

9. muellerberndt/cadence

⭐ 440 stars | Python

A brain that learns from live experience. System 1 by default: perception, memory, plasticity and action. Optional System 2 for self-observation and slow corrections.

brain-inspired embodied-ai equilibrium-networks local-learning machine-learning online-learning

10. TrenTorch/TrenTorch

⭐ 382 stars | Python

Don't memorize ML. Understand it from first principles including Inference, Cuda, Deep Learning, Data Science and Reinforcement Learning

cli cuda deep-learning education engineering from-scratch

11. Code-Amadeus/Amadeus

⭐ 371 stars | Python

Real-time multimodal desktop agent evolving toward a persistent AI OS interface (0.16 α).

ai-agent auip cuda desktop-agent desktop-companion electron

12. jrz97619761/test-model-thing

⭐ 325 stars | Python

A small proof-of-concept language model (not an LLM) that aims to incorporate latent-space prediction, internal state using recurrent trace units, and byte-by-byte output, built with MLX.

ai apple-silicon artificial-intelligence bytes continual-learning jepa

13. Speedstu/CUDA-for-AMD-Windows

⭐ 317 stars | PowerShell

Run CUDA-targeted Windows applications on AMD GPUs with ZLUDA + ROCm/HIP.

amd amd-gpu compatibility-layer cuda cuda-on-amd gpgpu

14. surya-koritala/Glyd

⭐ 248 stars | Rust

Lossless AI compression: 33% less GPU memory, bit for bit.

benchmark c-abi cli compression cuda data-compression

15. Akun-python/autumn-recruitment

⭐ 238 stars | Jupyter Notebook

Autumn-Recruit Algo Notes(秋招算法学习仓库)，面向大模型时代的算法、工程与AI学习共建仓库。从数学基础、数据结构，到机器学习、深度学习、LLM、Agent、AI-Infra、语音与数据科学，通过「理解→推导→手写 → 运行 → 验证 → 复盘构建可持续迭代的知识体系。

ai-agent ai-agents ai-infrastructure algorithms algorithms-and-data-structures autumn-recruit

Generated by Research Pulse on 2026-10-12 09:05